The architecture of massive parallel model fittings with GPU-LMFit.

<p>In pixel-wise parametric imaging applications, a large amount of experimental data for all image pixels are first collected and saved in host computer memory and then passed to the GPU global memory for efficient parallel model fitting analyses with GPU-LMFit using NVIDIA CUDA, which has a generic C-like language interface and it offers a data parallel programming model that is supported on NVIDIA GPUs. In CUDA, the host program launches kernel functions running on the GPU. A kernel function is organized as a hierarchy of threads. GPU Threads are grouped into blocks, and blocks are grouped into a grid. For each fitting, the parallel algorithms in GPU-LMFit, user-defined fitting function and/or Jacobian function are implemented in a single CUDA block by more than tens of GPU threads, which can synchronize and share data with each other and are all executed on a single GPU multiprocessor (MP). A GPU MP allows multiple GPU blocks running concurrently and a GPU device can have a number of GPU MPs, so using GPU-LMFit, a large number of LM model fittings can execute concurrently in a GPU, and each of the fittings is performed with parallel algorithms by many GPU threads in parallel.